Combining Confidence Elicitation and Sample-based Methods for Uncertainty Quantification in Misinformation Mitigation

TL;DR

This paper tackles misinformation mitigation with large language models by addressing the critical problem of unreliable uncertainty estimates due to hallucinations and overconfidence.It introduces a hybrid uncertainty framework that fuses verbalized confidence prompts with sample-based consistency, and evaluates a suite of prompting strategies, sampling methods, and truth scales on the LIAR dataset.Key contributions include a systematic comparison of sample-based methods, demonstration of the BSDetector framework in this domain, and identification of SampleAvgDev combined with a two-step elicitation as a particularly effective calibration approach, achieving an observed ECE as low as $0.076$.The findings underscore the value of hybrid uncertainty estimation for safer, more reliable misinformation mitigation and suggest generalizability to other high-stakes NLP tasks.

Abstract

Large Language Models have emerged as prime candidates to tackle misinformation mitigation. However, existing approaches struggle with hallucinations and overconfident predictions. We propose an uncertainty quantification framework that leverages both direct confidence elicitation and sampled-based consistency methods to provide better calibration for NLP misinformation mitigation solutions. We first investigate the calibration of sample-based consistency methods that exploit distinct features of consistency across sample sizes and stochastic levels. Next, we evaluate the performance and distributional shift of a robust numeric verbalization prompt across single vs. two-step confidence elicitation procedure. We also compare the performance of the same prompt with different versions of GPT and different numerical scales. Finally, we combine the sample-based consistency and verbalized methods to propose a hybrid framework that yields a better uncertainty estimation for GPT models. Overall, our work proposes novel uncertainty quantification methods that will improve the reliability of Large Language Models in misinformation mitigation applications.

Figures (11)

• Figure 1: Effect of Sample size on sample-based consistency methods
• Figure 2: Temperature ablation experiment on sample-based consistency methods
• Figure 3: BSDetector Framework
• Figure 4: Calibration curve for BSDetectork on SampleAvgDev
• Figure 5: Distributional Shift